Skip to content
aiqbenchAI & Quantum Intelligence
NewsInsightsBenchmarksSkillsAsk AILearnAbout
...Loading your session
NewsInsightsBenchmarksSkillsAsk AILearnAbout
Back to list
aiqbench takeChips & Infra

aiqbench analysis

The article covers software enablement for NVIDIA's Vera Rubin NVL72 rack, mainly vLLM inference support, Rubin-tuned FlashInfer kernels, and locality-aware MoE. Its headline 7.8x per-GPU throughput gain over GB200 NVL72 comes from a single benchmark (AgentX) and is an early, partly vendor-influenced result, so it should be weighed against independent benchmarks. For readers, the more durable signal is that Blackwell kernels carry over to Rubin and that inference-serving software is being co-designed with new hardware from day zero.
1
Source coveringvLLM Blog

vLLM Support for NVIDIA Vera Rubin NVL72: 7.8x Throughput over GB200 NVL72

vLLM Blog·October 9, 2026·10 min read

From vLLM Blog

Excerpt

vLLM now runs on NVIDIA Vera Rubin NVL72 with day-0 model support, Rubin-tuned FlashInfer kernels, and locality-aware MoE, delivering 7.8x the per-GPU throughput of GB200 NVL72 on AgentX.

vLLM now supports Vera Rubin NVL72!

NVIDIA Vera Rubin is the next-generation platform built for agentic inference. Inferact, NVIDIA, Red Hat, and the vLLM community have been bringing vLLM up on Vera Rubin NVL72 since it was announced, and vLLM runs on Vera Rubin NVL72 today with daily container builds and support for models from DeepSeek, Moonshot AI, Z.ai, and MiniMax.

Vera Rubin NVL72 hardware: 5x the NVFP4 FLOPS, about 2.4x the HBM bandwidth and 1.7x the bidirectional NVLink bandwidth of GB200 NVL72, with 2-4x faster exponentials for softmax.

Day-0 support: Rubin builds on Blackwell's architecture family, so vLLM’s Blackwell kernels are compatible with Rubin. Thanks to this, vLLM already supports diverse models such as DeepSeek, Kimi, GLM, and MiniMax on Rubin.

Rubin-tuned kernels: Through FlashInfer 0.7.0, vLLM gets Rubin-tuned attention, GEMM, and MoE kernels. We have also tuned our MiniMax Sparse Attention (MSA) prefill kernel for Rubin.

Locality-aware MoE: To make best use of Rubin's increased HBM bandwidth, we leverage CUDA 13.4's locality domains to split the MoE weights. This allows the SMs to read weights only from the memory nearest to them.

Excerpt shown under the publisher's RSS terms · 184 words · full text and figures are on the publisher's site

Read at vLLM Blog ↗Ask about this article

Related coverage

Chips & Infra

Niagara Networks launches 90 Gbps TLS decryption with post-quantum crypto

Quantum Zeitgeist · Nov 9, 2026

AI Research

Sparse Bayesian Modeling of EEG Channel Interactions Improves P300 Brain-Computer Interface Performance

arXiv · Sep 17, 2026

Machine Learning

DeltaVisionX: A Low-Cost Vision-Guided Automation System Using YOLOv11 and PLC-Based Actuation

arXiv · Nov 16, 2026

Chips & Infra

[AINews] Anthropic successfully completes largest IPO in history, valued at $1.5T

Latent Space · Nov 20, 2026

Article info

Source

vLLM Blog

Published

October 9, 2026

Read time

10 min

Category

Chips & Infra
aiqbench
AI & Quantum Intelligence© aiqbench. All rights reserved.

Feeds:RSS·JSON Feed

LanguageEnglish·Românăbeta
NewsAI & TechQuantumInsights
ExploreBenchmarksLearn
CompanyAboutTrustLegal noticeTerms of ServicePrivacy PolicyCopyright noticesData erasureReport content

The aiqbench take, the insights, the daily digest and Ask answers are written by AI models from the sources shown, without editorial review before publication. They may contain errors — check the source. How the model is used →

  • News
  • Insights
  • Ask AI
  • Benchmarks
  • Account